In 2026, AI agents can reliably take over the administrative half of running a private practice: scheduling, intake, appointment reminders, recall outreach, and drafting replies to patient messages. They cannot take over the clinical half. No agent should choose a diagnosis, pick a medication, or sign anything on your behalf. A doctor's license and the liability that comes with it set that boundary, regardless of how capable the technology gets.
What can an agent actually do today?
| An agent can handle | Stays with the doctor |
|---|---|
| Scheduling and calendar management | Diagnosis |
| Intake forms before a visit | Choosing or changing a medication |
| Appointment reminders | Signing a visit note |
| Recall and follow-up outreach | Approving a clinical order |
| Drafting patient message replies | Final say on any patient-facing action, even routine ones |
The left column is genuinely automatable because none of it requires clinical judgment. Booking a slot, sending a reminder, drafting a reply for review: these are administrative tasks that happen to touch patient data. The right column requires a license, and a license comes with a signature. A well-built agent can draft the note. It cannot sign it.
Where does this fit into an actual week?
A solo practice's admin load is bigger than most doctors want to think about, because most of it is invisible until it's late. Here's roughly where an agent takes weight off, and where it doesn't:
| Task | How often | What the agent does |
|---|---|---|
| Scheduling requests (book, reschedule, cancel) | Daily | Drafts the reply, proposes the calendar change for approval |
| Intake forms | Before each new visit | Sends the form ahead of the visit |
| Appointment reminders | Every visit | Sends automatically, 48 hours out and day-of |
| Recall and follow-up outreach | Ongoing | Sends automatically once the doctor sets the policy |
| Patient message replies (SMS, email) | Daily | Drafts a reply for the doctor or staff to send |
| Refill requests | Daily or weekly | Proposes the refill order, doctor approves |
| Superbill requests | As needed | Generates on request |
Reminders and recall outreach are the two rows that run without a human touching each instance, because the doctor already made the judgment call once, when setting the policy: send a reminder 48 hours before the visit and again the day of, recall a patient once enough time has passed since their last one. The agent just executes that policy on schedule. Everything else in the table either produces a draft for a human to send or a proposal for a human to approve.
Three ways to actually wire this up
There are three architectures in use right now, and they carry very different tradeoffs.
Copy-paste into a chatbot. A doctor pastes a patient's message into ChatGPT or Claude, gets a draft reply, and pastes it back into their EHR or texting app by hand. This works, and it's the lowest-effort way to start. It's also tedious at any real patient volume, and it creates a PHI handling problem: patient data is leaving your system of record and landing in a chat window that may or may not have a BAA behind it, with no record of what was pasted where. See using Claude as a doctor for how to do this as safely as the format allows.
Overlay vendors. A separate product connects to your existing EHR, usually through its API or by automating the browser, and writes actions back into it. These can add real capability to an EHR that has none of its own. The catch is that the overlay is a second system with its own permissions, its own audit trail, and its own understanding of your patients, bolted onto an EHR that wasn't built with an agent in mind. Every write-back is a translation, and translations are where bugs live.
An EHR with a native agent surface. The EHR itself exposes a set of tools an agent can call directly, scoped to that doctor's own data, with the same permission model the EHR already enforces for the doctor logged in through the browser. This is our approach at Eureka: the EHR itself is the agent substrate. We run a first-party MCP server, and any MCP-speaking client, Claude, Codex, Cursor, ChatGPT, can connect to it with a scoped API key. Competing "agent" products in this category are mostly overlays that write back into legacy EHRs, the pattern from option two. Other EHR vendors are moving toward native agent surfaces of their own, at different paces. See how EHR vendors are building MCP servers for a rundown of who has shipped what by mid-2026.
How does the doctor stay the decision-maker?
When the assistant in your practice's Slack channel wants to reschedule an appointment, cancel one, or send a refill order, it posts the proposed action and waits for a human. One click approves it, and nothing executes without that click. The doctor reads what's about to happen, in plain language, before it happens.
This matches how clinicians actually talk about AI in their own communities. On the psychiatry and therapist forums where doctors compare notes on tools, the line is consistent: automating the admin work around a practice draws support, and automating the judgment calls inside it draws pushback. A proposal-and-approval loop is what that line looks like in software. The agent handles the busywork of drafting and queuing. The doctor handles the one thing only a doctor can do, which is decide.
What actually changes once it's set up?
Setup itself is a short, one-time step: connecting an AI tool to your practice data, or turning on the assistant already built into your EHR. From there, the split holds steady.
Reminders and recall outreach run on their own from day one, since the doctor already made the judgment call once, when setting the policy, and no per-instance approval is required. Everything else, patient reply drafts, reschedule and cancellation proposals, refill orders, comes back to you or your staff for a read and a click, every time. The click gets faster as the drafts prove themselves, but it doesn't go away.
The parts that don't change, now or ever, are diagnosis, prescribing, and the note you sign, by design.
What about HIPAA?
Eureka signs a BAA with each practice we work with, covering the data that flows through our own platform. If a doctor connects a separate AI tool, Claude, ChatGPT, or anything else, directly to patient data, that practice is the covered entity and is responsible for having the right agreement in place with that AI vendor. Most consumer AI subscriptions don't include a BAA. Some vendors offer one on enterprise or API tiers, and both Anthropic and OpenAI launched healthcare-specific offerings in January 2026 that are worth checking against your own setup. Confirm BAA status with your AI vendor before sending any PHI through it. This isn't legal advice. One risk reducer worth building in regardless of vendor: allow only read tools for anything that processes patient-authored text, so a compromised or manipulated agent can see data but can't act on it.
More common questions about running a practice this way are answered in the FAQ.
To see it running in your own practice, start at eureka.md/signup.